Hacker News Reader

Comment by jashkenas | Hacker News Reader

Does anyone have a JavaScript implementation of the paper? It looks like it would be really fun to play with in <canvas>.

Edit Quoth YouTube: "74 minutes on an nVidia GeForce GTX 460" ... maybe not so fun.

View on HN

Here you go: http://jsfiddle.net/CSyUb/

Good job. Any idea why using the parameters specified in the video description of https://www.youtube.com/watch?v=KJe9H6qS82I do not seem to replicate the output of the video? http://jsfiddle.net/CSyUb/45/

I tried the same parameters in the version downloadable at http://sourceforge.net/projects/smoothlife/ and it did not seem to produce similar output to the video either.

They are using smooth time integration, where I am using the discrete scheme described in the arxiv paper. The switch isn't that difficult though, since it just involves modifying a single line of the shader.

The SmoothLifeConfig.txt doesn't have a line for SmoothLifeL - it should, Stephan!

Here it is:

2 1 10.0 3.0 10.0 0.100 0.257 0.336 0.365 0.549 2 4 4 0.028 0.147 // SmoothLifeL

Put this at the top of SmoothLifeConfig.txt and run the exe.

You'll notice that some of the parameters (the 2 4 4) control the transition function. This is different to the version in the paper which is smoothglider (which is in the config file) which uses 4 4 4.

Another thing about the 74 minutes. That was for a larger radius than necessary and a smaller timestep than necessary - for a high quality video.

With a radius of 10.0 and a timestep of 0.1 I get 186fps on the same graphics card - visually about the same speed as the youtube video.

But you missed the most important part of that quote. >"(in Ready, Stephan's software at the sourceforge link above runs much faster)" So using Ready it took a long time, but the actual code is much faster. I'd imagine you could write something in CUDA that would run this (and only this, not something generalized like Ready) plenty fast to do real time rendering on a new system, especially considering a GTX 680 has almost 5x the horsepower of a GTX 460.

Admittedly not in a web browser though.

Ready is using OpenCL. But Stephan is using FFT for the convolution with a disk filter, which is a big win for large radii.

As far as I'm aware OpenCL performance, while portable, is still somewhat inferior to that of straight CUDA on NVIDIA GPUs, hence my CUDA suggestion. I didn't actually look at the algorithms used in either so I can't comment, perhaps I'll do that tomorrow. What do you mean by using an FFT "for" a convolution? The algorithm is convolving the (presumably 2D) FFT of the game board with a disk filter?

SmoothLife uses two disks around each point to determine the instantaneous rate of change. Inside each disk the values are summed. This is a convolution with a disk filter.

http://en.wikipedia.org/wiki/Convolution

As it says on that page, FFT is often used for convolution because it is fast: after applying a discrete Fourier transform to the kernel and the image, the resulting images must only be multiplied together before applying an inverse FFT.

http://en.wikipedia.org/wiki/Discrete_Fourier_transform

That's so cool. I love to see real world applications of convolutions.

I wonder if there is an application of Gershgorin's Disc theorem here?

I see something about a circle that contains the eigenvectors of a matrix. How does that apply here?

Ah, thanks for the information. I'm familiar with signal processing, so the only way I'm used to seeing convolutions is through the multiplication of FFTs. I wasn't even considering the "regular" way.

"Admittedly not in a web browser though."

HTML5 gives you WebGL support, which allows you to compile and run GLSL shaders from JavaScript. For something that is heavily shader-oriented like this, it might be possible in the browser.

WebCL?

Reply on news.ycombinator.com